Toward trustworthy ML on decentralized frameworks: fairness in learning and feedback in communication

Abstract

As machine learning (ML) becomes prevalent in our daily lives, it is now employed to make critical decisions that affect our lives, cultures, and rights, e.g., screening job applicants and informing bail & parole decisions. With a surge of such applications, one major criterion in the design of ML systems is trustworthiness. To this end, there has been increasing interest in developing an ML system that is not only accurate but also fair and privacy-preserving. In this dissertation, we make efforts toward trustworthy ML on decentralized frameworks. Specifically, we make progress on how we ensure fairness in machine learning, as well as how we enable a higher data rate via feedback for efficient decentralized learning in wireless networks.

In the first part, we focus on fairness in machine learning. Fairness, which aims to design a model that does not discriminate among different demographics and/or individuals, is one key aspect for enabling trustworthy ML. First, we study fairness issues that occur in the context of classifiers. We develop a kernel density estimation (KDE) methodology for a fair classifier that comes with a high accuracy-fairness tradeoff. Second, we explore the fairness issue that arises in recommender systems. Biased data due to inherent stereotypes of particular groups (e.g., male students' average rating on mathematics is often higher than that on humanities, and vice versa for females) may yield a limited scope of suggested items to a certain group of users. We introduce a novel fairness notion (that we call equal experience), which can serve to regulate such unfairness in the presence of biased data. We also propose an optimization framework that incorporates the fairness notion as a regularization term, as well as introduce computationally-efficient algorithms that solve the optimization.

In the second part, we study the fundamental role of feedback in wireless communication. Privacy-preserving machine learning is another key aspect of trustworthy ML. To this end, decentralized learning has been employed instead of collecting all data which contains the private information of users. This decentralized training requires a significant amount of information exchange in wireless networks during the training process. Therefore, wireless impairments such as interference and noise significantly affect the training performance. So as an initial effort to develop an efficient decentralized machine learning framework, we study interaction via feedback in wireless networks to improve the capacity performance of networks. In this dissertation, we explore a deterministic two-way IC which captures key properties of the wireless Gaussian channel, and completely characterize the capacity region of this channel (w.r.t. the forward and backward sum-rate pair) via a new achievable scheme and a new converse. One surprising consequence of this result is that not only we can get an interaction gain over the one-way non-feedback capacities, but we can also sometimes get all the way to perfect feedback capacities in both directions simultaneously.